# Fresh environment record (A6, 2026-11-10)

Built from a fresh clone of git branch pherc0332-repro-s1_f327 (61562810a) with `cd spiral-fitting && uv sync --frozen`
in a new directory. The original workstation venv was not used.

- Host: Ubuntu 26.04, Linux 7.0.0, GCC 15.2.0, CMake 4.2.3, uv-selected CPython 3.14.4 (/usr/bin/python3.14).
- torch 2.11.0+cu128 (CUDA 12.8). The GPU was an NVIDIA GeForce RTX 4070 on driver 580.178.04. The re-scoring is CPU only.
- Native prerequisites: `cmake_configure.log` (the configure output of the vc_spiral and vc_object_store builds).
  The list in ../README.md is a summary of that log.
- `pip_freeze_fresh.txt`: the fresh environment. `pip_freeze.diff`: compared with the original venv (../../eval/venv.json):
  - editable paths differ (expected);
  - the fresh env lists vc_object_store as an editable install, and the original did not list it;
  - py-spy (a profiler) was only in the original;
  - vc-delta3d is 0.1.0 in the original vs 0.2.0 from the lock (sidecar packing only; see ../README.md).
- Stock-host build, 2026-11-10: `minimal_host/` (clean Ubuntu 26.04 + only the Dockerfile's apt packages) PASSED. See `minimal_host.log`
  and `uv_sync_minimal.log`. `pip_freeze_minimal.txt` equals `pip_freeze_fresh.txt` except for the editable paths. The first attempt with
  `libopencv-core-dev` failed because that package has no `OpenCVConfig.cmake` (`minimal_host_attempt1_opencv.log`); `libopencv-dev` is required.

## Acceptance A6b: exact re-scoring without workstation paths
`a6b_sandbox_score.sh` ran `run_reference.sh score` inside bubblewrap. In that sandbox:
- `/media` was absent, and of /home only the fresh clone and env were present (checked inside the sandbox);
- the data was a copy of the published files;
- the source was the published frozen_source tar.

Result (`a6b_score.log`): #81 s1, #81 s2, M6 s1 and M6 s2 all reproduce the published Vb `scores_trim.json`. In each,
all 4,139 leaves are equal (4,011 numeric, plus the checkpoint, model and manifest hashes), with the same types and
key structure. Only the 5 relocatable path fields differ.

## Arm assembly
- `assemble_dataset_tta_hub1030h_trim.log` (#81) and `assemble_dataset_tta_hub1030h_trim_aircrop_m1108.log` (M8): the
  arms were rebuilt from the published files, archive members and delta with 0 differences. #81 was verified on path, type,
  sha256, size and mtime. The M8 log is from before mtimes were restored, so it was verified on everything but mtime.
- `assemble_dataset_tta_hub1030h_trim_aircrop_m1106.log` (M6): rebuilt with the final assembler, 0 differences on path,
  type, sha256, size and mtime; `track_caches_dataset_tta_hub1030h_trim_aircrop_m1106.log`: all its track caches accepted.
- `track_caches_dataset_tta_hub1030h_trim.log`: on the rebuilt #81 arm, the frozen track loader accepts both DBMs'
  packed stores and crossing caches.
